What was originally a joke, a colloquial term for projects that are done on the weekend, in which you tell a model what you want to do, don't read what it tells you, and follow it by feeling. It was completely out of that "hole" by 2026. Ninety-two per cent of U.S. developers are using AI coding tools daily, up 340% from the year before, and a growing number of junior developers built their entire portfolio around it, since it allowed them to deliver something impressive without having to go through college and years of fundamental coding.
That remained for a time. Next, it was time for the bill.
The surging demand for hiring that no one anticipated.
Vibe coding isn't only a coding style; it's an economic change. 25% of startups in Y Combinator's Winter 2025 batch had codebases that are 95% or more AI-generated. The price of developing a viable SaaS product has come down from $200,000 to $5,000, and the time to build is now six weeks versus six months. In that same context, a junior developer who can truthfully boast, "I built this whole app myself using AI" seems like a brilliant hire, both cheap and fast and able to deliver a complete product with just one developer.
63% of today's coding users describe themselves as non-developers, product managers, marketing directors, founders, designers. Forrester reckons that there are currently more than 16 million active citizen developers in the world, and Gartner foresees the number of citizen developers will surpass the number of professional engineers by four to one by 2028. For a hot minute, it was really just "junior developer" or "person who could prompt well", and hiring managers went for whoever could get the demo packaged out the quickest.
Next, the codebases had to be tested!
All this has a productivity understory that isn't as easy as the pitch makes it sound. In a controlled study, METR determined that more experienced developers who worked on large, actual codebases were 19% slower when using AI tools, and also felt they were 20% faster. Almost 40 points between how fast people feel and how fast they actually are, and that's important in that it really means that the teams that are planning sprints based on their self-reported productivity are doing it quietly on the sidelines.
The same can be said for quality. In one prominent study, the rate of bugs increased by 41% post AI coding tool adoption, while other research comparing AI-coauthored pull requests with those written by humans revealed 1.7 times more major issues, with logic errors up 75% and readability issues tripling. However, the less glamorous part of code, that which is refactored and simplified, dropped from a quarter of changed code in 2021 to under ten per cent by 2024, even as the amount of code being shipped kept climbing. This doesn't appear on day one. It emerged by the 12th week, when teams are usually using 20-30% of their sprints to repair bugs they weren't fully aware of when they were first created by AI code.
That's the reason why junior developers are the first to take the hit. Still, debugging doesn't involve writing code; it involves understanding code, and that's something that vibe coding allows you to avoid. A developer can develop a working feature and never, ever have a mental model of why it works, and that's okay until it doesn't and there's no foundation under them. Add security to the mix, and you'll find that almost half of all code generated by AI fails basic industry security standards, and that one study of more than 200 applications coded by vibe reveals that over 90 per cent had one or more vulnerabilities stemming from a lack of awareness of what they don't know. These are not fancy bugs, but the very basic ones: unvalidated input, lack of access controls, exposure of credentials, all of that sort of thing that an experienced engineer instinctively looks for, and a junior developer who didn't get the rudiments of engineering basics.
There is no uncertainty about the production failures. One vibe-coding tool experienced three separate security incidents which were widely covered, including one that remained open for 48 days. In another case, an AI-powered social app took just three days to leak 1.5 million authentication tokens and 35,000 emails. On the broader market, about ten thousand startups have tried to work out a working production application using an AI assistant, and over eight thousand of them need costly rebuilds that are usually $50,000 to $500,000 a pop, because the demo they would build wasn't going to hold up with actual users, actual load, and an actual security audit.
The market took no notice of this,
As developers have incorporated these tools into their workflows, they are also becoming more sceptical about their effectiveness. Developers' own trust in the tools has been shaken as they have used them more. Trust has dropped from approximately 40% two years ago to approximately 29% today, and almost half of developers actively distrust the output of AI-generated code. Only a small number of senior engineers say they have high trust, and one in five say they have high distrust among senior engineers specifically. But less than half of developers claim to always look over AI-generated code before committing it, meaning that plenty of AI-generated code is flying the flag without a second glance.
Labour statistics are now catching up with the anecdotes. Overall demand for engineering talent hasn't gone away, but junior developer hiring has dropped by almost 20% among those who would typically begin their careers in this role. As much as forty to seventy per cent of the traditional entry-level development has been swallowed up by AI. Meanwhile, junior developers, who are seeking to enter the field, are doing so in one of the most competitive fields in years, and at the bare entry-level, in a field that increasingly demands the judgment of a senior engineer.
The irony is sharp. The idea of vibe coding was to make it the leveller, the one thing that anyone could do and be a builder. For a certain group of junior coders, however, it was a way to an appointment and a quick path out of it, since the skills used to land that job didn't necessarily help when it came time to interact with real people, real traffic, real security audits.
What really makes the difference between the ones who make it and those that don't?
The organisations and developers that will emerge from this cycle without issues are not ones that eschewed AI but rather embraced it as an aid to real understanding, not a replacement. The pattern that passes this test is more "let me build with AI" than "let AI build it for me," with the caveat that while the AI is allowed to build it, there is a human in charge of checking each step it takes when the code is written into production code, the AI doesn't run the test gates unless there is a deadline pressure that can't be overcome, and the specifications are detailed enough for an agent to know exactly what is supposed to be built, and not just handed an open-ended prompt and encouraged to guess well. Such teams are consistently spared the technical debt and exposure to security issues that plague other teams that scale up their use of AI without first establishing that structure. If teams consider governance as something to do afterwards, they are almost always going to pay for it afterwards at multiple times the cost of doing it from the beginning.
Vibe coding doesn't need to be abandoned, and it shouldn't. If you're a junior developer trying to land a job right now, the pitch that lands you the job and keeps you on is NOT "I can have AI create things." It's "I know how it works well enough to fix it when it breaks, secure it before it ships, and explain it to the next person who touches it."